HAL Id: hal-00397403
https://hal.archives-ouvertes.fr/hal-00397403v2
Preprint submitted on 5 Jul 2010
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires
A note on choosing the alternative with the best median evaluation
Jean-François Laslier
To cite this version:
Jean-François Laslier. A note on choosing the alternative with the best median evaluation. 2010.
�hal-00397403v2�
A NOTE ON CHOOSING THE ALTERNATIVE WITH THE BEST MEDIAN EVALUATION
Jean-François LASLIER
Cahier n° 2010-17
ECOLE POLYTECHNIQUE
CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE
DEPARTEMENT D'ECONOMIE
Route de Saclay 91128 PALAISEAU CEDEX
(33) 1 69333033
http://www.economie.polytechnique.edu/
mailto:chantal.poujouly@polytechnique.edu
A note on choosing the alternative with the best median evaluation
Jean-François Laslier
Département d’Économie, École Polytechnique 91128 Palaiseau, France
Jean-Francois.Laslier@polytechnique.edu First version May 2009, this version June 2010.
Abstract
The voting rule proposed by Basset and Persky (Public Choice 99: 299- 310) picks the alternative with best median evaluation. The present note shows that this MaxMed principle is equivalent to ask the social planner to apply the MaxMin principle allowing him to discard half of the population.
In one-dimensional, single-peaked domains, the paper compares this rule with majority rule and the utilitarian criterion. The MaxMed outcome is rejected by a majority of voters in favor of outcomes which are also utilitarian improvements.
1 Introduction
When di¤erent individuals evaluate one object on some common scale, taking as a summary of the various evaluations their median is a natural idea, often used in Statistics. Using this idea for collective choice among several alternative objects is thus occasionally proposed: the suggestion is that an alternative with the largest median evaluation is to be chosen. I will call this method the MaxMed voting scheme.
In the modern literature this was (up to my knowledge) proposed by Basset
and Persky, 1999 [2]. Properties of the best median can be expressed in the
language of utility theory (Bossert and Weymark 2004 [3]): the informational
basis is the ordinal and inter-individuallly comparable framework. In terms
of utilities, the utility levels attached by the di¤erent individuals to a given
alternative are compared (inter-individual comparability), and the median is
stable under any strictly increasing transformation of utility provided that the
transformation is the same for all individuals (ordinality). This is the same
informational basis as for the MaxMin, or Rawls principle. As will be shown,
instead of maximizing the satisfaction of the least favored individual in the
whole society, as the MaxMin does, MaxMed maximizes the satisfaction of the
least favored individual when half of the population can be discarded.
As to strategic aspects, supporters of this method claim that it is relatively immune to individual misrepresentation of evaluations. The reason for this claim is that the median (contrary to the mean) is “robust” in the statistical sense.
For instance, in most cases, over-evaluating an object in order to push up its median has simply no consequence. This argument was put forward by Basset and Persky [2], who used the term “Robust Voting” to describe this method.
But, in fact, voting rules generally share the property that one voter is rarely pivotal so that, in practice, median voting does not appear to be less, nor more, manipulable than other voting rules (Gerhlein and Lepelley 2003 [6]).
Some properties of the MaxMed method are collected by Warren Smith at the Center for Range Voting [10]. Using a limited set of grades, several candi- dates usually end up with the same grade. Balinski and Laraki, 2007 [1] have proposed several, more or less complicated, ways to choose among them, and have ellaborated on the question of robustness to manipulation. Felsenthal and Machover, 2008 [4] is a discussion of these issues.
In this paper I will leave aside the strategic questions and assume that voters vote sincerely. I also leave aside the problem of ties associated with the use of MaxMed voting in practice, by assuming that an evaluation can be any real number. I will attempt to compare the outcomes of di¤erent choice principles, including MaxMed, in a setting which is standard in Economic and Political Theory and is relevant for the applications. I consider one-dimensional, single- peaked utility pro…les with distributions of voters’ideal points which are skewed (without loss of generality: skewed to the left).
From the point of view of Social Choice Theory, one-dimensional, single peaked pro…les are very speci…c pro…les because they avoid Condorcet cycles, but they nevertheless constitute an interesting benchmark case. In Political Economy, this assumption is so common that it is often not even mentioned.
The idea that the distribution of voters with respect to the relevant parameter is not uniform or symmetric but skewed is an empirical observation: it seems to be a general rule that socioeconomic relevant parameters are “skewed to the left,” the paradigmatic example being income distributions: most people earn less than the average.
In this setting, Condorcet-consistent voting rules are very simple, and all alike: all of them chose the median of the voters’ideal points, which is a Con- dorcet winner in virtue of the celebrated Median Voter Theorem. This is the outcome of majority voting. By comparison, the utilitarian choice (the e¢ cient alternative in the sense of maximizing the sum of individual utilities, I will call it the Bentham winner) tends, in the same setting, to produce choice which are favored by richer people. Although this observation is not as clear-cut as the Median Voter Theorem is, it matches the economic intuition and can be stated formally if one makes the (standard) assumption of quadratic utility. In that case the utilitarian optimum is simply the mean of the distribution of ideal points. For left-skewed distribution, the mean is larger than the median.
Then, where is the alternative with the best median evaluation located? Is
it close to the utilitarian optimum or not, and if not, in which direction does it
diverges from the optimum: to the left, in the direction of the ideal points of
the majority, or to the right, in the direction of the rich minority? How does it compare to the outcome of Majority rule, the Condorcet winner? Is it more or less e¢ cient than this outcome?
In order to answer this question, I present one very simple analytical exam- ple, plus computer simulations under various hypothesis which will be described in the sequel. The reached conclusion is always the same: the best median is located on the wrong side of the Condorcet winner. The Bentham-inne¢ ciency of the best median choice is of the same kind but worse than the Bentham- inne¢ ciency of majority voting.
The example maybe useful in order to understand what one does when com- paring medians. The usual argument for rejecting an alternative A in favor of another alternative B when a majority of individuals prefer A to B is that members of the relatively small population who gain in this move gain a lot while members of the losing majority incur a relatively small loss. This is a typ- ical Benthamite, utilitarian argument. Conversely, the Bentham-inne¢ ciency of majority voting derives from the democratic power of the (many) poor
1. This ine¢ ciency can be justi…ed by normative political arguments in favor of the prin- ciple of majority rule. It can also often be justi…ed by invoking the argument of decreasing individual marginal utility. In the one-dimensional settings under scrutiny, the best median choice goes further away from e¢ ciency, for reasons that are easily understood from the mathematical point of view (the example in section 5 will explain this point) but which have no normative or political appeal: apply Rawls principle to the most homogeneous half-population, with no regard for the other half.
2 Interpretation of the MaxMed
In order to explain what one does when chosing the alternative with the best possible median evaluation, it is useful to introduce some de…nitions and pieces of notation.
De…nition 1 Consider a population P of individuals and a set X of alterna- tives. The Rawlsian evaluation of x 2 X by P is:
R(P; x) = min
i2P
u
i(x) and the Rawlsian satisfaction of P is:
R(P ) = max
x2X
R(P; x):
Notice that this vocabulary may not be exactly true to John Rawls’ ideas (Rawls 1971 [8]), but it is customary in Economics to call Rawls’principle the idea of giving full priority to the worst-o¤ individuals (Kolm 1972 [7], Fleurbaey and Hammond 2004 [5]).
1
At the limit one can wonder why, under majority rule, “the poor do not expropriate the
rich”; see Roemer, 1998 [9].
De…nition 2 Let I, with measure 1, be the (in…nite) set of inviduals and let H be the set of all sub-populations of measure 1=2. Suppose that, for an alternative x, the evaluation u
i(x) of x by i is continuously distributed in I. Then we denote by med
i2Iu
i(x) the median evaluation of x, uniquely de…ned by the fact that for some half-population H 2 H ,
u
j(x) med
i2Iu
i(x) u
h(x) for all h 2 H and all j 2 I n H .
Notice that for a …nite population, if the number of individual is odd then there are no subpopulation of measure 1=2, and that if the number of individuals is odd then the median is generally not unique For simplicity I consider from now an in…nite population in which the median evaluation of each alternative is well de…ned, and I shall indicate how the …ndings extend to the …nite case.
Let x
MaxMedbe an alternative with the largest median evaluation and let u
MaxMedbe this evaluation. By de…nition those are obtained by the following algorithm:
1. For each alternative x 2 X, compute u
med(x) = med
iu
i(x) the median of the evaluations of x by all individuals.
2. Find the largest of these numbers u
MaxMed= max
xu
med(x) and an associ- ated alternative x
MaxMedwhich achieves this maximum.
The next lemma will show that the following de…nition is equivalent:
1. For each half-population H of P , compute its Rawlsian satisfaction R(H ).
2. Find the largest of these numbers u
MaxMed= max
HR(H ) and an associated alternative x
MaxMedwhich achieves this maximum.
Lemma 1
max
x2X
med
i2I
u
i(x) = max
H2H
R(H )
Proof. Write u = max
x2Xmed
i2Iu
i(x) and r = max
H2HR(H). To see that u r , take any x 2 X . By de…nition of the median there exists a half-population H 2 H such that 8 j 2 H, med
i2Iu
i(x) u
j(x). For this H , R(H; x) = min
j2Hu
j(x) med
i2Iu
i(x), which implies that R(H ) med
i2Iu
i(x). Thus r med
i2Iu
i(x), hence r u .
Conversely, suppose u > r . Then there exists " > 0 and there exists x 2 X such that med
i2Iu
i(x) > r + ". For that x:
8 H 2 H ; med
i2Iu
i(x) > R(H) + ":
By de…nition of the median there exists H
0such that u
h(x) med
i2Iu
i(x) for all h 2 H
0, hence, for this H
0, u
h(x) > R(H
0) + ". But R(H
0) = max
y2XR(H
0; y) thus, for y = x, we have that
8 h 2 H
0; u
h> R(H
0; x) + ":
Taking the min on h, one obtains the contradiction R(H
0; x) R(H
0; x) +
": Q.E.D.
For a …nite population of odd size n, the same result holds when considering for H the subpopulations of size (n + 1)=2. Here is, for illustration, a simple example with three voters and two alternatives:
u
i(x) i = 1 i = 2 i = 3 med
iu
i(x)
x = a 6 7 8 7
x = b 4 3 9 4
Table 1: best median evaluation.
R(H; x) H = f 1; 2 g H = f 1; 3 g H = f 2; 3 g
x = a 6 6 7
x = b 3 4 3
R(H ) 6 6 7
Table 2: Most favored half population
3 The one-dimensional model
The set of available alternatives is the set of positive real numbers X = [0; + 1 [.
To each individual i is attached an ideal point x
i2 X . The utility of i for an alternative y 2 X is denoted u
i(y). Under the quadratic utility assumption one has:
u
i(y) = (y x
i)
2: (Variants of this assumption will be considered later.)
The society is then described by the distribution of ideal points. Let F denotes the cumulative distribution: for any x 2 X, F (x) is the proportion of individuals i such that x
i< x. We suppose that F is continuous and has a mean.
The Condorcet winner is then well-de…ned and unique, it is the median of the distribution, that is the alternative, denoted x
Condsuch that:
F (x
Cond) = 1=2:
The alternative which maximizes the total utility W (y) = R
X
(y x)
2dF (x) is the average of the ideal points. We call this point the Bentham optimum and denote it by x
Bentham:
x
Bentham= Z
x
x dF(x):
For any alternative y, the utility levels obtained by the various individuals
induce a probability distribution that we denote by E
y: for any utility level
u 0, E
y(u) is the proportion of individuals in the society whose utility for
y is less than u. Given our assumption about utilities, these individuals are precisely those whose ideal points are at a distance larger than p
u from y:
u
i(y) < u () (y x
i)
2< u () x
i2 = y p
u; y + p u thus
E
y(u) = 1 F (y + p
u) + F(y p u)
and the median evaluation of y, denoted u
med(y) is such that E
y(u
med(y)) = 1=2, that is:
F(y + p
u
med(y)) F (y p
u
med(y)) = 1=2:
It may be more convenient to write this formula with d(y) = p
u
med(y) as:
Z
y+d(y) y d(y)dF (x) = 1=2:
The best median evaluation choice, the outcome of “median evaluation”
is denoted by x
MaxMed, its is the point which maximizes u
medor, equivalently, which minimize d. In general it is di¢ cult to compute this point, but the example in the next sections makes these computations very simple.
4 Two results for non-symetric distributions
Here we consider the case of continuous distribution of ideal points with a density that is continuous and monotonous on its support. Without loss of generality we take the density to be decreasing and the support to be S = [0; b] or S = [0; + 1 ).
Theorem 1 Suppose that the distribution F of voters’ ideal points has a con- tinuous density f on its support S and that f is decreasing on S. Suppose that voters have quadratic utilities u
i(y) = (y x
i)
2. Then x
Benthamx
Cond. Proof. The fact that x
Benthamx
Condis just the fact that, for such a dis- tribution, the average is larger than the median. For a proof, consider G the distribution made of a mass m at 0 plus a uniform distribution on [0; c].
Routine computation shows that, for G, the median is smaller than the av- erage. Take G(x) =
12+ (x x
Cond) f(x
Cond), c = x
Cond+
2f(x1Cond)
and m = 1=2 x
Condf (x
Cond). The median of G is x
Cond. The cumulative func- tion G is above F so F stochastically dominates G and thus the average of F is larger than the average of G. Therefore the average of F is larger than its
median. Q.E.D.
Theorem 2 Suppose that the distribution F of voters’ideal points has a contin- uous density f on its support S and that f is decreasing on S. Suppose that vot- ers have utilities ordinally equivalent to quadratic utilities u
i(y) = (y x
i)
2for strictly increasing. Then x
MaxMed=
12x
Cond.
Proof. Note that for any population P with support [y; y + 2d], the Rawlsian satisfaction is obtained for the choice y + d and has value R(P ) = ( d
2=4).
Therefore, in view of Lemma 1, x
MaxMedis the value of y + d that minimizes d under the constraint that Z
y+2dy
dF = 1=2
for some y 0. This constraint can be written as F (y + 2d) = 1=2 + F (y) or, equivalently:
d = (1=2)F
1(1=2 + F(y)) (y):
Given our hypotheses, the function ( ) is diferentiable for y 0, and 2
0(y) = f (y)
f (y + 2 (y)) 1:
Because f is decreasing, the ratio in this expression is smaller than 1, thus is decreasing too, and is maximized for the smallest value y = 0. It follows that F(2d) = 1=2, which means that x
Cond= 2d. Hence the result. Q.E.D.
5 A simple example
Suppose that the distribution of voters’ideal point has bounded support [0; 1]
with an a¢ ne density function parametrized by a skewness parameter , with 0 < 1:
f (t) = 1 + 2 t:
For = 0 the density is uniform on [0; 1] and for > 0 the density is skewed on the left. For = 1 the density is triangular. Figure 1 is drawn for = 1=3.
Then for 0 x 1, F (x) =
Z
x 0(1 + 2 t) dt = (1 + )x x
2:
Majority Voting: Condorcet winner. See point C in Figure 1. The median of the ideal points is such that F (x
Cond) = 1=2, that is 2 x
22(1 + )x + 1 = 0, which gives:
x
Cond= 1
2 1 + p
1 +
2:
x
Benthamx
Condx
MaxMed0 1=2 1=2 1=4
= 1=2 :42 :38 :19
= 1 :33 :29 :15
Table 1: Choices for an a¢ ne distribution with skewness
Some values are provided in Table 1. When the skewness tends to 0, x
Condtends to 1=2.
Utilitarian evaluation: Bentham optimum. See point B in Figure 1. The average of the ideal points is x
Bentham= R
20
tf (t)dt and computation shows:
x
Bentham= 1 2 6 :
Again, when the skewness tends to 0, x
Benthamtends to 1=2, like x
Cond. Maximal Median evaluation: See point M in Figure 1. In order to compute the median evaluation at a point y one has to …nd a radius d(y) such that half of the ideal points are located in the segment [y d(y); y + d(y)], and half are located outside. The best median is then obtained at the point y such that d(y) is the smallest. In Figure 1, one can see that the best median is precisely the mid-point between 0 and the Condorcet winner x
Cond. Indeed, by de…nition of x
Cond, half of the population belongs to the segment [0; x
Cond], and because the density is decreasing, any other segment of length smaller or equal will contain strictly less than half of the population. Therefore:
x
MaxMed= x
Cond2 = 1
4 1 + p
1 +
2:
When the skewness tends to 0, x
MaxMedtends to 1=4, unlike the two other choices.
One can see on this simple example the result announced in the introduction:
Proposition 1 For any a¢ ne distribution f of skewness > 0:
x
MaxMed< x
Cond< x
Bentham:
Proof. This follows from the theorems above, but a direct proof is easy: For any 2 (0; 1], the inequality
211 + p
1 +
2<
12 6is shown algebraically by letting u = p
1 +
2, and the result follows. Q.E.D.
Notice that the phenomenon, which holds for any value of the skewness
parameter , is all the most striking for very small vallues of , which coresponds
to distributions which are almost not skewed. In that case, letting tend to 0, the
B
0 M C 1
Figure 1: Three choices from a skewed distribution of ideal points
Condorcet winner and the utilitarian optimum both tend to 1=2, the center of the distribution: for almost symmetric distributions, there is no con‡ict between the utilitarian and the majoritarian principles: both chose a consensual outcome in the middle of the distribution. But things are totally di¤erent with MaxMed, whose choice tends to 1=4. As we have seen this choice is due to the fact that the MaxMed calculus allows to leave aside half of the population.
The graphical representation makes transparent the explanation of the phe-
nomenon. Half of the population lies on each side of the Condorcet winner
x
Cond= C, but the right side (between C and 1) is wider than the left one (be-
tween 0 and C). This should be an argument in favor of a collective choice larger
than C, like the utilitarian optimum x
Bentham= B, because individuals at the
left of C are relatively close to C while those at the right of C are relatively far
from C. This is a typical utilitarian argument, that weights numbers of individ-
uals and intensity of preferences. The reasoning of the MaxMed argument, that
leads to the choice x
MaxMed= M , is reversed: the window being narrower on
the left and wider on the right, chosing a point in the middle of the left window
will satisfy the left half of the population and will reach for the — already quite
satis…ed — members of this group a relatively high level of satisfaction because
this group is not too diverse. On the contrary, choosing a point on the right of
M , such as C or B, the satis…ed half of the population would be spread over a
larger segment and it will be more di¢ cult to reach the same level of satisfaction
for this half population, because these people are more diverse.
- 1 1 2 3 4 5 0.1 0.2 0.3 0.4 0.5 0.6
Figure 2: Log-Normal distribution
The choice of M is dictated by the level of satisfaction obtained by some half of the population. By de…nition of the MaxMed, this level is the largest than can be obtained by any half of the population, thanks to inter-individual comparability. But, doing so, it neglects the (maybe very low) level of satisfac- tion obtained by the other half of the population, thanks to ordinality. There is no compromise here: …nd the half of the population which would be better o¤ if they were in power ! This “majoritarian” logic is ‡awed, as one can see on the example, because the result is that moves away from M in the direction of the center simultaneously (i) satisfy the majority criterion: more losers than winners, and (ii) satisfy the utilitarian criterion: the losers loose less than what the winners win.
6 Simulated examples with Log-Normal distri- bution
In the simulations, I compute the Bentham optimum, the outcome of Majority Rule (the Concorcet winner), the outcome of the Borda rule (the Borda winner) and the point with the best median evaluation. I work with a Log-Normal distri- bution, which is typically the kind of distribution met in Social Sciences (Figure 2) rather than a¢ ne distributions as in the previous section. The theoretical mean of the Log-Normal distribution of parameters 0 and 1 is 1:649 and the stan- dard deviation is 2:161. The distribution vanishes quickly after x = 5 and, when needed, I restrict attention to the segment [0; 5].
2I pick at random 999 ideal points according to this distribution. Then, in a …rst example, the individual
2
This has no impact on the value of the Condorcet winner nor on the MaxMed. It leads to
a slight underestimation of the Bentham optimum and the Borda winner.
candidates n
ku
i(y) Bentham Borda Condorcet MV
d
22 1 1 :5
11 d 1 1 1 :5
uniform d
1=21 1 1 :5
d
21:77 1:15 1:04 :62
49 d 1:04 1:15 1:04 :62
d
1=2:83 1:15 1:04 :62
d
21:68 :92 :92 :56
11 d :92 :92 :92 :56
repres. d
1=2:83 :92 :92 :56
d
21:89 :93 1:04 :56
49 d 1:04 :93 1:04 :56
d
1=2:85 :93 1:04 :56
Table 2: Various speci…cations and choices with a Log-Normal Distribution
utility functions are quadratic: u
i(y) = (y x
i)
2, and there are 11 candidates evenly spread between 0 and 5 (at points 0, :5, 1, 1:5, 2; 2:5, :::, 5). In that case, the Bentham optimum is located at x
Bentham= 2, the Borda winner and the Condorcet winner are both located at x
Cond= x
Borda= 1, and the best median evaluation is obtained at 0:5, thus: x
MaxMed< x
Cond= x
Borda< x
BenthamAn alternative speci…cation is that the utility is decreasing linearly with the distance: u
i(y) = j y x
ij . I label this case linear utility. Going from one speci…cation to the other does not change the Condorcet or Borda winner nor does it change the median evaluations, because this is a strictly increasing trans- formation common to all individuals. But it changes the utilitarian optimum, which is now equal to the Condorcet winner. (Recall that the solution to the problem min
yP
i